Hoof, Herke van (2016)
Machine Learning through Exploration for Perception-Driven Robotics.
Technische Universität Darmstadt
Dissertation, Erstveröffentlichung

Kurzbeschreibung (Abstract)

The ability of robots to perform tasks in human environments has largely been limited to rather simple and specific tasks, such as lawn mowing and vacuum cleaning. As such, current robots are far away from the robot butlers, assistants, and housekeepers that are depicted in science fiction movies. Part of this gap can be explained by the fact that human environments are hugely varied, complex and unstructured. For example, the homes that a domestic robot might end up in are hugely varied. Since every home has a different layout with different objects and furniture, it is impossible for a human designer to anticipate all challenges a robot might face, and equip the robot a priori with all the necessary perceptual and manipulation skills.

Instead, robots could be programmed in a way that allows them to adapt to any environment that they are in. In that case, the robot designer would not need to precisely anticipate such environments. The ability to adapt can be provided by robot learning techniques, which can be applied to learn skills for perception and manipulation. Many of the current robot learning techniques, however, rely on human supervisors to provide annotations or demonstrations, and to fine-tuning the methods parameters and heuristics. As such, it can require a significant amount of human time investment to make a robot perform a task in a novel environment, even if statistical learning techniques are used.

In this thesis, I focus on another way of obtaining the data a robot needs to learn about the environment and how to successfully perform skills in it. By exploring the environment using its own sensors and actuators, rather than passively waiting for annotations or demonstrations, a robot can obtain this data by itself. I investigate multiple approaches that allow a robot to explore its environment autonomously, while trying to minimize the design effort required to deploy such algorithms in different situations.

First, I consider an unsupervised robot with minimal prior knowledge about its environment. It can only learn through observed sensory feedback obtained though interactive exploration of its environment. In a bottom-up, probabilistic approach, the robot tries to segment the objects in its environment through clustering with minimal prior knowledge. This clustering is based on static visual scene features and observed movement. Information theoretic principles are used to autonomously select actions that maximize the expected information gain, and thus learning speed. Our evaluations on a real robot system equipped with an on-board camera show that the proposed method handles noisy inputs better than previous methods, and that action selection according to the information gain criterion does increase the learning speed.

Often, however, the goal of a robot is not just to learn the structure of the environment, but to learn how to perform a task encoded by a reward signal. In addition to the weak feedback provided by reward signals, the robot has access to rich sensory data, that, even for simple tasks, is often non-linear and high-dimensional. Sensory data can be leveraged to learn a system model, but in high-dimensional sensory spaces this step often requires manually designing features. I propose a robot reinforcement learning algorithm with learned non-parametric models, value functions, and policies that can deal with high-dimensional state representations. As such, the proposed algorithm is well-suited to deal with high-dimensional signals such as camera images. To avoid that the robot converges prematurely to a sub-optimal solution, the information loss of policy updates is limited. This constraint makes sure the robot keeps exploring the effects of its behavior on the environment. The experiments show that the proposed non-parametric relative entropy policy search algorithm performs better than prior methods that either do not employ bounded updates, or that try to cover the state-space with general-purpose radial basis functions. Furthermore, the method is validated on a real-robot setup with high-dimensional camera image inputs.

One problem with typical exploration strategies is that the behavior is perturbed independently in each time step, for example through selecting a random action or random policy parameters. As such, the resulting exploration behavior might be incoherent. Incoherence causes inefficient random walk behavior, makes the system less robust, and causes wear and tear on the robot. A typical solution is to perturb the policy parameters directly, and use the same perturbation for an entire episode. However, this strategy tends to increase the number of episodes needed, since only a single perturbation can be evaluated per episode. I introduce a strategy that can make a more balanced trade-off between the advantages of these two approaches. The experiments show that intermediate trade-offs, rather than independent or episode-based exploration, is beneficial across different tasks and learning algorithms.

This thesis thus addresses how robots can learn autonomously by exploring the world through unsupervised learning and reinforcement learning. Throughout the thesis, new approaches and algorithms are introduced: a probabilistic interactive segmentation approach, the non-parametric relative entropy policy search algorithm, and a framework for generalized exploration. To allow the learning algorithms to be applied in different and unknown environments, the design effort and supervision required from human designers or users is minimized. These approaches and algorithms contribute towards the capability of robots to autonomously learn useful skills in human environments in a practical manner.

Frage zum Eintrag

Redaktionelle Details anzeigen